Embodiments use a multi-modal approach to generate software programs that match a solution program description. The solution program description may include natural language, input-output examples, partial source code, desired operators, or other hints. Some embodiments use optimized prompts to a pre-trained language model to obtain initial candidate programs. Maximal program components are extracted and then recombined variously using component-based synthesis. Beam search reduces a solution program search space by discarding some candidates from a given synthesis iteration. Relevance metrics, string similarity metrics, operator frequency distributions, token rareness scores, and other optimizations may be employed. By virtue of optimizations and the multi-modal approach, a solution program may be obtained after fewer iterations than by use of a language model alone. The multi-modal approach is domain agnostic, as illustrated by examples using regular expression and cascading style sheet selector domain specific languages.
Legal claims defining the scope of protection, as filed with the USPTO.
3. The computing system of claim 1, further comprising a synthesized programs cache containing at least one of the following: one or more of the obtained program components, or one or more of the constructed candidate programs.
4. The computing system of claim 1, further comprising semantic equivalence classes, each semantic equivalence class containing one or more respective program components, each program component in a given semantic equivalence class producing upon execution a same output in response to an input of an input-output pair of the solution program description.
7. The method of claim 6, wherein the method is domain-agnostic in that the method is parameterized by at least a domain specific language.
9. The method of claim 6, further comprising utilizing a maximal component as an initial program component while iteratively constructing the set of candidate programs, the maximal component having a higher frequency of occurrence in the obtained set of program components than at least one other component in the obtained set, the maximal component also having a larger size than at least one other component in the obtained set.
10. The method of claim 6, wherein iteratively constructing the set of candidate programs comprises either retaining or eliminating a given candidate program based on at least a beam search result.
11. The method of claim 6, wherein iteratively constructing the set of candidate programs comprises condensing programs by using semantic equivalence classes.
13. The method of claim 6, wherein ranking candidate programs comprises calculating a string similarity between a candidate program and a program component obtained from the language model.
14. The method of claim 6, further comprising identifying a closed program as a candidate program, wherein the closed program is a program which has a type or a sort that is a start symbol of a domain specific language grammar.
15. The method of claim 6, further comprising selecting a program component for use as a candidate program based on at least one of the following: a probability of occurrence of the program component, or a probability of redundancy of the program component.
17. The computer-readable storage device of claim 16, wherein obtaining the set of program components from the language model comprises preparing a prompt, submitting the prompt to the language model, and extracting a program component from a prompt completion provided by the language model responsive to the prompt, wherein the prompt includes a sample question-answer pair, and wherein the method further comprises choosing the sample question-answer pair for inclusion in the prompt at least in part by using a relevance metric, the relevance metric being based at least in part on a token rareness score.
18. The computer-readable storage device of claim 16, wherein ranking candidate programs comprises calculating a string similarity between a candidate program and a program component obtained from the language model.
19. The computer-readable storage device of claim 16, wherein a candidate program includes a sketch, the sketch being a component template which matches multiple components.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
December 7, 2021
March 19, 2024
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.